Qwen3.8-27B-pi-adapted for Apple Silicon (MLX, 8-bit)

harness-hub/Qwen3.8-27B-pi-adapted (revision e786f84) merged into Qwen/Qwen3.8-27B (revision 3756181) and converted to MLX with mlx-vlm, vision encoder and Qwen's MTP layer included. In this build the linear layers are quantized to 8-bit (affine, group 64). Size: 30.0 GB.

harness-hub trained this LoRA with reinforcement learning (AsyncGRPO in TRL) to be the model behind the pi coding agent on software-engineering and terminal tasks. Their published results, under pi with thinking off: Terminal-Bench 2.1 39.5% โ†’ 55.8% and SWE-bench Pro (hard 51) 29.4% โ†’ 62.7% against the base model, with 55-64% fewer tool calls per task. See the original adapter card for the method, data and full results.

Build Size Same next token as bf16 KL vs bf16
netprojects/Qwen3.8-27B-pi-adapted-MLX-6bit (recommended) 23.1 GB 99.0% 0.0015
netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit 30.0 GB 99.1% 0.0004
netprojects/Qwen3.8-27B-pi-adapted-MLX-4bit 16.3 GB 95.6% 0.0201
netprojects/Qwen3.8-27B-pi-adapted-MLX-bf16 55.6 GB reference reference
netprojects/Qwen3.8-27B-pi-adapted-MLX-adapter 638 MB LoRA only, for mlx-lm on any Qwen3.8-27B MLX build

The 6-bit build is the recommended balance of size and fidelity; 8-bit is closest to bf16, 4-bit is the smallest.

Use with pi

The adapter was trained and evaluated with thinking off, so keep it off. Serve the model with mlx-lm's OpenAI-compatible server:

pip install -U "mlx-lm>=0.32.2"
mlx_lm.server --model netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit --port 8000 \
  --chat-template-args '{"enable_thinking": false}'

Add it to ~/.pi/agent/models.json:

{
  "providers": {
    "local": {
      "baseUrl": "http://localhost:8000/v1",
      "api": "openai-completions",
      "apiKey": "dummy",
      "models": [{ "id": "netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit" }]
    }
  }
}

"api": "openai-completions" matters: the adapter was trained and evaluated with the Chat Completions API, not the Responses API. Then:

pi --provider local --model netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit --print "Find out why the tests in tests/test_parser.py fail and fix the bug."

Use from Python

Text only, with mlx-lm:

from mlx_lm import load, generate

model, tokenizer = load("netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit")
messages = [{"role": "user", "content": "Write a Python function that parses '1.5M' into bytes."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

With images, with mlx-vlm:

pip install -U "mlx-vlm>=0.7.11"
mlx_vlm.generate --model netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit --image screenshot.png \
  --prompt "What error does this terminal show?" --max-tokens 300   # thinking stays off unless --enable-thinking

The tokenizer and chat template are harness-hub's files, unchanged (same vocabulary and tool-call format as Qwen3.8-27B). Tool calls use Qwen's XML format (<tool_call><function=...><parameter=...>), which mlx-lm's server returns as OpenAI tool_calls.

Faster decoding: MTP and DSpine

This build keeps Qwen3.8-27B's own multi-token-prediction (MTP) layer in model-mtp.safetensors, quantized to match the build. The adapter does not change that layer, so it is Qwen's original MTP head. mlx-vlm uses it as a built-in draft model when you pass the same repository as --draft-model:

mlx_vlm.generate --model netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit --draft-model netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit \
  --draft-kind mtp --temperature 0 --max-tokens 512 --prompt "Write a Python function that checks whether a number is prime."

On this build MTP makes decoding about 2.2ร— faster with no extra download. The DSpine drafter with dspine-hybrid-mlx also works with these builds. Our measurements (greedy, thinking off, 5 prompts ร— 256 tokens from GSM8K, MATH-500, HumanEval, MBPP and MT-Bench, with other services running on the Mac, so treat them as indicative):

Build Plain MTP (mlx-vlm 0.7.11) DSpine v0.2.1 (4-bit drafter, block 5)
4-bit 33 tok/s 61 tok/s (ร—1.84) 70 tok/s (ร—2.25)
6-bit 23 tok/s 22 tok/s (ร—0.96) 52 tok/s (ร—2.32)
8-bit 18 tok/s 41 tok/s (ร—2.24) 54 tok/s (ร—3.05)

Both methods check every drafted token with the full model. With greedy decoding the text matches plain decoding until a near-tie, where the batched check can round the other way, so long answers can diverge slightly.

With Splash

These builds also run in Splash (1.7.0), which uses its own DFlash2 draft model for speculative decoding. Measured on the same 5 prompts ร— 256 tokens, thinking off, on 19 November 2026 with other models loaded on the Mac, so treat the speeds as indicative; text, tool calling and image input passed on every build:

Build Decode speed End to end, incl. prompt Tokens per cycle
4-bit 185 tok/s 135 tok/s 6.8
6-bit 135 tok/s 104 tok/s 6.8
8-bit 115 tok/s 90 tok/s 6.9
hf download netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit --local-dir Qwen3.8-27B-pi-adapted-MLX
splash serve --model-dir Qwen3.8-27B-pi-adapted-MLX --model local/qwen3.8-pi --reasoning-effort none

Splash loads the quantized builds (4, 6 and 8-bit); it does not load bf16 MLX checkpoints.

What was checked

On a MacBook Pro M5 Max (128 GB), 18 November 2026, with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11.

  • Merge. All 256 adapted modules (q, k, v, o in the 16 full-attention layers; gate, up, down in all 64 layers) equal W + 2.0 ยท B @ A rounded to bf16, checked element by element against an independent float64 computation; every other tensor is byte-identical to Qwen3.8-27B. The adapter does not touch the vision encoder or the Gated DeltaNet layers.
  • Quantization. Each quantized build is checked to be exactly the affine (group 64) quantization of the bf16 build.
  • Agreement with bf16 (12 short prompts, thinking off, 961 teacher-forced positions): the 8-bit build picks the same next token as bf16 on 99.1% of positions, 6-bit on 99.0% and 4-bit on 95.6% (mean KL 0.0004 / 0.0015 / 0.0201).
  • Tool calling with the included chat template, thinking off: one agent-style prompt with bash and read tools; 8-bit, 6-bit, 4-bit each produced a well-formed bash tool call in Qwen's XML format.
  • Coding. Three short Python tasks run against asserts: 8-bit 3/3, 6-bit 3/3, 4-bit 3/3.
  • Vision (mlx-vlm): asked to name the shapes and colours in a test image, 8-bit, 6-bit, 4-bit answered correctly, and answered a text-only question in the same session.
  • MTP layer. The bundled MTP tensors equal Qwen3.8-27B's mtp.* weights (norms in mlx-vlm's stored form), quantized exactly like the rest of the build; each build loads normally in mlx-lm and mlx-vlm with the extra file present.
  • Source files. Every file of both pinned upstream revisions matches the Hub's SHA-256 hashes.

Credits

The adapter is by harness-hub; the base model is Qwen3.8-27B by the Qwen Team, Alibaba Cloud. If you use the model, please cite their work (see the adapter card and the Qwen3.8-27B card).

License

Apache-2.0, as Qwen3.8-27B and the harness-hub adapter. LICENSE is the Apache-2.0 text from the Qwen3.8-27B repository and NOTICE describes the conversion and credits harness-hub and the Qwen Team (Alibaba Cloud).

Published by NetProjects (Internet Projects Ltd). Not affiliated with Hugging Face, harness-hub or Alibaba.

More MLX builds from NetProjects: huggingface.co/netprojects ยท netprojects.com

Downloads last month
-
Safetensors
Model size
28B params
Tensor type
U32
ยท
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit

Base model

Qwen/Qwen3.8-27B
Quantized
(16)
this model

Collection including netprojects/Qwen3.8-27B-pi-adapted-MLX-8bit